Time Series Classification Based on Image Transformation Using Feature Fusion Strategy

Time series classification is an important branch of data analysis. Scholars have proposed a large number of time series classification methods in recent years. However, time series classification remains a challenging problem due to feature selection in time series classification. In order to further simplify the feature selection procedure and improve time series classification accuracy, an automatic feature selection of a time series classification method based on an image feature fusion strategy and a deep learning algorithm is proposed. First, a time series is transformed into images using different types of image transformation methods, i.e. the recurrence plot, Gramian angle difference field, Gramian angle summation field and Markov transition field. Second, the above four images are encoded into a new image type, that is the combined image, by a feature fusion strategy. Finally, a convolutional neural network is used for combined image classification and forecasting model selection. Time series from the M1 and M3 competition datasets are used to verify the effectiveness of the proposed method. The experimental results show that the algorithm has a higher classification accuracy and smaller prediction error compared to the benchmark models. Moreover, the forecasting error MAPE of combined image method is reduced by 0.2020 and 1.7454 compared with the traditional image method and single forecasting methdd respectively.